Source-linked AI summary
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, Zhiyuan Liu
TL;DR
Text evaluation remains costly and automatic metrics often correlate weakly with human judgments, especially for open-ended or specialized tasks. ChatEval addresses this gap with autonomous multi-agent debate using diverse roles and communication strategies, and reports evaluation performance more closely aligned with human preferences.
Problem
Human annotation is time- and cost-intensive, while n-gram-based automatic metrics correlate weakly with human judgments for open-ended or domain-specific generation.
Method
ChatEval uses autonomous multi-agent debate, assigning agents distinct roles and communication strategies to collaboratively evaluate generated responses.
Results
ChatEval improves text-quality evaluation performance and aligns more closely with human preferences across open-ended questions and NLG tasks.
Takeaways & Limitations
Diverse role specification and distinct communication strategies are integral to ChatEval’s reliable, comprehensive evaluations akin to human judgment.
Takeaways & Limitations
The study uses homogeneous groups of LLMs from the same GPT family model, leaving heterogeneous groups for future research.
Abstract
from arXiv · showhide
Text evaluation has historically posed significant challenges, often demanding substantial labor and time cost. With the emergence of large language models (LLMs), researchers have explored LLMs' potential as alternatives for human evaluation. While these single-agent-based approaches show promise, experimental results suggest that further advancements are needed to bridge the gap between their current effectiveness and human-level evaluation quality. Recognizing that best practices of human evaluation processes often involve multiple human annotators collaborating in the evaluation, we resort to a multi-agent debate framework, moving beyond single-agent prompting strategies. The multi-agent-based approach enables a group of LLMs to synergize with an array of intelligent counterparts, harnessing their distinct capabilities and expertise to enhance efficiency and effectiveness in handling intricate tasks. In this paper, we construct a multi-agent referee team called ChatEval to autonomously discuss and evaluate the quality of generated responses from different models on open-ended questions and traditional natural language generation (NLG) tasks. Our analysis shows that ChatEval transcends mere textual scoring, offering a human-mimicking evaluation process for reliable assessments. Our code is available at https://github.com/chanchimin/ChatEval.
1 INTRODUCTION
Text evaluation is costly, while traditional automatic metrics often correlate weakly with human judgments on open-ended or specialized tasks. ChatEval addresses this gap by using debating LLM agents with distinct roles and communication strategies.
- Human annotation is time- and cost-intensive, while n-gram metrics often correlate weakly with judgments for open-ended or domain-specific generation.
- LLM-as-a-judge methods use instruction-following language models to evaluate open-ended responses and traditional NLG tasks.
- ChatEval models collaborative human evaluation by having multiple LLM agents discuss judgments while adopting distinct personas and expertise.
- ChatEval is proposed as a multi-agent framework intended to align better with human preferences than single-agent approaches.
- The framework contributes communication strategies and diverse role prompts for multi-agent debate, alongside a composable and scalable library.
2 METHODOLOGY
ChatEval organizes autonomous LLM debaters with distinct role prompts and configurable communication strategies. Final judgments use aggregation procedures suited to comparison or direct-scoring formats.
- Debater Agents: ChatEval treats each LLM as an autonomous debater that receives other agents’ responses as chat history and requires no human intervention.
- Diverse Role Specification: Distinct role prompts assign different personalities to agents while preserving a shared prompt template.
- Communication Strategy: ChatEval defines communication strategies as alternative ways to maintain and manipulate agents’ chat histories.
- One-By-One: One-by-one communication lets agents respond sequentially using preceding agents’ messages in their chat histories.
- Simultaneous-Talk: Simultaneous-talk has agents generate responses asynchronously to reduce the impact of speaking order.
- Simultaneous-Talk-with-Summarizer: Simultaneous-talk-with-summarizer adds an LLM that summarizes each debate iteration for all agents’ chat histories.
- Final Aggregation: For direct comparisons ChatEval uses majority voting, whereas direct scores are aggregated by averaging annotator scores.
3 EXPERIMENTS
ChatEval is evaluated on open-ended question answering and dialogue response generation using human-judgment-based benchmarks. Results compare multi-agent evaluation with single-agent, traditional, and other LLM-based methods.
- Benchmarks: ChatEval is evaluated on FairEval for open-ended question answering and Topical-Chat for dialogue response generation.
- Implementation Details: The experiments use GPT-4 or ChatGPT agents at temperature 0, with homogeneous groups in which all agents use the same GPT-family model.
- Open-Ended Question Answers: FairEval contains 80 open-ended questions, with human reference judgments derived from majority votes among three annotators.
- Dialogue Response Generation: Topical-Chat evaluates responses across naturalness, coherence, engagingness, and groundedness using annotations for 60 dialogue contexts.
- Evaluation Metrics: Accuracy and Kappa measure classification correctness and agreement with human annotators while accounting for chance agreement.
- Dialogue Response Generation: ChatEval improves GPT-4 dialogue-evaluation performance over G-EVAL, increasing average Spearman correlation by 0.096 (16.3%).
4 ANALYSIS
ChatEval analysis examines diverse roles, communication strategies, role counts, discussion turns, and qualitative debate behavior on FairEval. Diverse roles and one-by-one communication improve evaluation, while additional discussion turns show no consistent benefit.
- ChatEval analysis compares diverse role prompts, communication strategies, role numbers, discussion turns, and qualitative debate behavior using FairEval and ChatGPT.
- 4.1 THE IMPORTANCE OF DIVERSE ROLE PROMPTS: Diverse role prompts outperform identical prompts and can improve evaluation beyond the single-agent setting.The identical-role configuration underperforms the diverse-role configuration and does not effectively enhance performance over a single agent.
- 4.2 THE STUDY OF COMMUNICATION STRATEGIES: One-by-one communication is more effective than the other tested strategies, although all three exceed the naive single-agent method.The experiments use three agents, two discussion turns, and diverse role prompts.
- 4.3 THE IMPACT OF ROLE NUMBERS AND DISCUSSION TURNS: 62.5% accuracy is reached with 3 or 4 roles on FairEval before declining at 5 roles, while agreement continues increasing from 3 to 4 roles.Accuracy and agreement rise with role number up to the reported peak, but additional roles do not uniformly improve accuracy.
- 4.3 THE IMPACT OF ROLE NUMBERS AND DISCUSSION TURNS: Increasing discussion turns produces no significant upward trend and may stagnate or degrade performance as context length grows.The passage attributes this pattern to issues associated with ever-increasing context length.
- 4.4 QUALITATIVE ANALYSIS: In a stress-advice example, debate ends with both responses judged equally good, matching human annotations after agents weigh detail, concision, and context.The agents move from differing preferences to a consensus that the responses have the same quality.
- 4.4 QUALITATIVE ANALYSIS: The qualitative process includes opening a stance, proposing alternatives, and reaching consensus through mutual understanding and compromise.
- 4.4 QUALITATIVE ANALYSIS: ChatEval presents evaluation as interactive dialogue rather than static single opinions, capturing nuances through collaborative discussion.
5 RELATED WORK
Related work covers automatic NLG evaluation, communicative agents, and ChatEval’s experimental examples and analyses. The cited material contrasts reference-based metrics and multi-agent interaction while illustrating debate configurations and response judgments.
- Automatic NLG evaluation: Automatic NLG evaluation addresses labor-intensive human annotation through computational assessment of generated text.
- Automatic NLG evaluation: N-gram metrics such as ROUGE and BLEU are widely used but cannot capture syntactic and semantic similarity.
- Communicative Agents: Communicative agents use natural-language interaction and negotiation to support more productive collaboration on complex tasks.
- ChatEval examples: ChatEval’s example compares two comprehensive stress-management responses whose differences center on concision, detail, and additional techniques.
- ChatEval analysis: Figure 3 evaluates role-count and discussion-turn effects under one-by-one communication, using two turns for role comparisons and two roles for turn comparisons.
- Communicative Agents: Prior Peer Rank and Discussion work uses different models as agents to probe evaluation dimensions but does not explore alternative communication strategies.
6 CONCLUSION
ChatEval improves text-quality evaluation performance while aligning more closely with human preferences. Its diverse roles and communication strategies support comprehensive, human-like evaluation discussions.
- ChatEval improves text-quality evaluation performance and aligns more closely with human preferences.
- The framework treats diverse role specifications and communication strategies as integral components of multi-agent evaluation.
- Qualitative discussion analysis provides insight into ChatEval’s evaluation process and supports its reliability and efficacy.
A PROMPT TEMPLATE AND DIVERSE ROLE PROMPT
ChatEval uses diverse referee roles in its prompt template, assigning agents distinct perspectives for evaluating generated responses. These roles collectively support critical, article-consistent, behavioral, and scientifically grounded judgments.
- The prompt template assigns referees distinct role descriptions to diversify evaluation perspectives.
- General Public focuses on the story and investigation updates while independently selecting the better response.
- Critic checks fluent writing, clear sentences, and wording, while questioning judgments and proposing alternatives when responses are equally strong.
- News Author evaluates consistency with the original article and helps determine which response is better.
- Psychologist considers human behavior and mental processes to help determine which response is better.
- Scientist contributes systematic study, scientific-method, critical-thinking, and problem-solving perspectives to response evaluation.
B FORMAL DEPICTION OF DIFFERENT COMMUNICATION STRATEGY
ChatEval formalizes communication strategies for multiple debate agents that exchange responses through shared chat histories or buffers. The process extracts final evaluation results using majority voting or average scores.
- The FairEval prompt template is also used for single-agent experiments, with chat history and role-description slots ignored.
- The one-by-one strategy passes each agent’s current response to later agents and adds a summarized buffer to every agent’s chat history.
- The formal strategies take agent number, discussion turns, debate agents, and chat histories as inputs, then produce final text-evaluation results.
- The simultaneous-talk strategy collects current-turn responses in a buffer and shares that buffer with all agents.
- Final results are extracted using either majority vote or average score.